Method for constructing density functional tight constraint-neural network hybrid potential function

Abstract

The invention relates to the technical field of material science, and discloses a method for constructing a density functional tight constraint-neural network hybrid potential function, and the method comprises the steps: obtaining structure parameters of a target material sample in a data set through a constructed density functional theory data set; calculating energy, atomization energy and stress data of an electronic structure corresponding to the sample configuration based on a density functional theory; calculating band structure energy and stress of a target material sample in the density functional theory data set based on a density functional tight constraint theory, and obtaining corresponding density functional tight constraint atomization energy; calculating repulsion energy and repulsion stress of each sample configuration according to each atomization energy and the corresponding stress data; and constructing an atomic neural network potential function, and combining the atomic neural network potential function with a density functional tight constraint theory to generate a hybrid potential function. According to the method, the DFTB method and the machine learning technology are combined, the precision and the mobility of the potential function are greatly improved, and meanwhile, the calculation efficiency is remarkably improved.

Description

Technical Field

[0001] The invention relates to the field of material science and technology, and in particular to a method for constructing a density functional tight binding-neural network hybrid potential function. Background Art

[0002] The ab initio molecular dynamics (AIMD) method based on density functional theory (DFT) is widely used in simulation and prediction of materials science due to its high accuracy. However, the AIMD method has a very high computational cost, especially for the calculation of large-scale material systems, which consumes huge computing time and resources. Especially when dealing with systems containing hundreds of atoms, long-term simulation becomes extremely difficult. On the other hand, although semi-empirical methods and force field methods perform well in simulation efficiency, they have low accuracy and poor portability. This accuracy and portability problem limits its application in complex material systems.

[0003] Density Functional Tight Binding (DFTB) theory is a semi-empirical electronic structure calculation method based on DFT. Its calculation speed is 2-3 orders of magnitude faster than that of DFT method, but its accuracy is limited. In the traditional DFTB framework, the parameterization of repulsive energy is the most complicated step, which is essentially a complex multidimensional fitting problem between atomic pair potential and DFT reference results. The limitation of this paired form limits the accuracy and transferability of DFTB. With the continuous development of machine learning technology and the continuous accumulation of first-principles calculation data, machine learning (ML) as an emerging tool has achieved many successes in the field of computational catalysis. Using neural network (NN) potential function to replace traditional DFT methods can greatly improve the calculation efficiency. It has been implemented in multiple software packages, such as SchNetPack, DeePMD-Kit, LASP, etc. In terms of accuracy, the data-driven NN potential can approach the accuracy of the provided DFT training set within a certain chemical space. The root mean square error of the NN potential function reported in most literatures is 5-20meV / atom for energy and 20meV / atom for force. The machine learning potential function requires that all possible application scenarios be covered as much as possible in the training data. However, during the MD simulation process, a large number of unknown species configurations such as reaction intermediates and free radicals will be generated, and the system conformational space will be very large and difficult to predict. This has also become the main bottleneck for the use of NN potential functions for MD simulation predictions.

[0004] Combining physical models with machine learning to construct potential functions can improve the transferability of potential functions, especially improve the description of a large number of unknown species configurations that may be generated by high-temperature collisions outside the training set; on the other hand, compared with the full machine learning method, it can achieve the required accuracy in a smaller training set. Therefore, it is urgent to construct an efficient and accurate DFTB-NN potential function as a method to solve the problem of MD calculation efficiency and accuracy, so as to achieve accurate and efficient molecular dynamics simulations. Summary of the invention

[0005] In view of this, the present invention provides a method for constructing a density functional tight binding-neural network hybrid potential function to solve the problems of MD calculation efficiency and accuracy, so as to achieve accurate and efficient molecular dynamics simulation.

[0006] The present invention provides a method for constructing a density functional tight binding-neural network hybrid potential function, the method comprising: constructing a density functional theory data set, obtaining structural parameters of a target material sample in the data set, and calculating the energy, atomization energy and force data of the electronic structure corresponding to the sample configuration based on the density functional theory; calculating the band structure energy and force of the target material sample in the density functional theory data set based on the density functional tight binding theory, and obtaining the corresponding density functional tight binding atomization energy; calculating the repulsion energy and repulsion force of each sample configuration based on the atomization energy and force data calculated by the density functional theory, and the atomization energy and force data of the corresponding sample configuration calculated by the density functional tight binding; constructing an atomic neural network potential function based on the above calculation results, and using the repulsion energy and repulsion force as labels to train the neural network potential function; combining the trained neural network potential function with the density functional tight binding theory to generate a hybrid potential function, performing error analysis on the hybrid potential function, and optimizing the hybrid potential function according to the analysis result.

[0007] In an optional embodiment, before calculating the energy, atomization energy and force data of the electronic structure corresponding to the sample configuration based on density functional theory, it also includes: obtaining the sum of the energies of each atom and the total molecular energy of the constituent molecules in the target material sample; based on the sum of the energies of each atom and the total molecular energy of the constituent molecules, calculating the atomization energy of each sample configuration.

[0008] In an optional embodiment, the formula for calculating the atomization energy of each sample configuration of the target material sample is:

[0009] E at =E sum_atoms -E molecule

[0010] Among them, E at is the atomization energy of the sample configuration, E sum_atomsis the sum of the energies of the atoms that make up the molecule, E molecule is the total energy of the molecule. In an optional embodiment, the repulsion energy and repulsion force of each sample configuration are calculated using the following formula:

[0011] E rep =E at_DFTB_el -E at_DFT

[0012] F rep =F DFTB_el -F DFT

[0013] Among them, E at_DFTB_el The band structure atomization energy, E, obtained by density functional tight binding calculation for each sample configuration at_DFT is the atomization energy calculated by density functional theory for each sample configuration, F DFTB_el The density functional tight binding force data for each sample configuration, F DFT Density functional theory force data for each sample configuration.

[0014] In an optional embodiment, an error analysis is performed on the hybrid potential function, specifically: describing the repulsive energy and repulsive force based on the neural network potential function; combining the density functional tight binding band structure energy and the neural network repulsive energy, and combining the potential function force with the neural network repulsive force; and performing an error analysis on the hybrid potential function based on the combination result.

[0015] In an optional embodiment, the use of repulsive energy and repulsive force as labels to train the neural network potential function includes: obtaining a density functional tight binding data set based on density functional tight binding calculation results; dividing the density functional tight binding data set into a training set, a validation set and a test set; using the repulsive energy and repulsive force in the training set as labels to train the neural network potential function, the validation set is used to adjust the sample, and the test set is used to evaluate the performance of the trained neural network potential function.

[0016] In an optional embodiment, after dividing the density functional tight binding data set into a training set, a validation set and a test set, it also includes: creating a data loader, reading the training set, the validation set and the test set into the memory of the data loader, and preprocessing them to obtain a data form that can be used by the neural network potential function.

[0017] In an optional embodiment, an atomic neural network potential function is constructed based on the above calculation results. The multi-body local interactions in the density functional tight binding data set can be fitted using neural network potential function fitting software including but not limited to SchNetPack, DeePMD-Kit, LASP, etc. to obtain the neural network potential function.

[0018] Compared with the prior art, the technical solution provided by the embodiments of the present disclosure has the following advantages:

[0019] The present invention provides a method for constructing a density functional tight binding-neural network hybrid potential function, which comprises: constructing a density functional theory data set, obtaining structural parameters of a target material sample in the data set, and calculating energy and force data of an electronic structure corresponding to the target material sample; obtaining the band structure energy and force of the target material sample in the data set based on the density functional tight binding theory, and calculating the atomization energy of the corresponding density functional tight binding; calculating the repulsion energy and repulsion force of each sample configuration based on the atomization energy and force calculated by the density functional, and the atomization energy and force of the corresponding sample configuration calculated by the density functional tight binding; constructing an atomic neural network potential function based on the above calculation results, and using the repulsion energy and repulsion force as labels to train the neural network potential function; combining the trained neural network potential function with the density functional tight binding theory to generate a hybrid potential function, performing error analysis on the hybrid potential function, and optimizing the hybrid potential function according to the analysis result. The present invention combines the DFTB method and machine learning technology, greatly improving the accuracy and transferability of the potential function. At the same time, the present invention only calculates the band structure energy part in DFTB, which significantly improves the calculation efficiency. It also decomposes the energy force of the system, one part is calculated based on DFTB, and the exclusion part is fitted through a neural network, which makes up for the problem that the traditional DFTB method is insufficient in describing multidimensional problems in complex structures, and reduces the dependence of machine learning on specific data sets, and has wide applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 is a schematic diagram of a process for constructing a density functional tight binding-neural network hybrid potential function according to an embodiment of the present invention;

[0022] Figure 2is a schematic diagram of a process for constructing a DFTB-NN hybrid potential function using a small molecule system as an example according to an embodiment of the present invention;

[0023] Figure 3 It is a comparison of the accuracy of the DFTB-NN potential function and the NN potential function on the QM9 data set using a small molecule system as an example according to an embodiment of the present invention;

[0024] Figure 4 It is a comparison of the accuracy of the DFTB-NN potential function and the NN potential function in the FeCHO system taking the iron-based catalytic system as an example according to an embodiment of the present invention;

[0025] Figure 5 It is a comparison of the accuracy of the DFTB-NN potential function, the DFTB potential function and the NN potential function in the FeC cluster system taking the iron-based catalytic system as an example according to the embodiment of the present invention; DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0027] Density functional theory (DFT) is a method for studying the electronic structure of multi-electron systems. Although it provides high-precision calculation results, its calculation cost is very high. The present invention can significantly reduce the calculation time and resource consumption and greatly improve the research efficiency by using the combination of density functional tight binding (DFTB) method and neural network. In practical applications, the method of the present invention can be integrated into software tools such as DFTB+ software, SchNetPack, DeePMD-Kit, LASP and ASE, without large-scale software and hardware transformation, which is further convenient for relevant researchers to use.

[0028] According to an embodiment of the present invention, a method embodiment for constructing a density functional tight binding-neural network hybrid potential function is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0029] In this embodiment, a method for constructing a density functional tight binding-neural network hybrid potential function is provided. Figure 1 is a flow chart of a method for constructing a density functional tight binding-neural network hybrid potential function according to an embodiment of the present invention, such as Figure 1 As shown, the process includes the following steps:

[0030] Step S100, constructing a density functional theory data set, obtaining the structural parameters of the target material sample in the data set, and calculating the energy, atomization energy and force data of the electronic structure corresponding to the sample configuration based on the density functional theory.

[0031] In this embodiment, detailed information of the target material, such as elemental composition information, molecular quantity, 3D structure and other information, is obtained, and the above detailed information is constructed into a density functional theory data set based on density functional theory (DFT). At the same time, the spatial coordinate information of the target material is obtained according to the contained 3D structure information, and the energy and force data of the electronic structure corresponding to the target material sample are calculated using the DFT method.

[0032] Step S200, calculating the band structure energy and the force of the target material sample in the data set based on the density functional tight binding theory, and obtaining the corresponding density functional tight binding atomization energy.

[0033] In this embodiment, the DFTB+ software is used to calculate the band-structure energy and the stress of each sample configuration in the target material sample, and the repulsive term (ie, the repulsive term) is not calculated.

[0034] In this embodiment, the electronic structure energy and force data corresponding to the target material sample are calculated based on DFT, and the band structure energy and force of each sample configuration are calculated based on DFTB. Before calculating its atomization energy, it also includes: obtaining the sum of the energies of each atom and the total molecular energy of the constituent molecules in the target material sample; calculating the atomization energy of each sample configuration based on the sum of the energies of each atom and the total molecular energy of the constituent molecules.

[0035] In this embodiment, the single-atom energy of each molecule in the target material sample is obtained by using the DFTB+ software, and the single-atom energies are aggregated to obtain the sum of the energies of each atom and the total energy of the molecule.

[0036] In this embodiment, the formula for the atomization energy of each sample configuration of the target material sample is:

[0037] E at =E sum_atoms -E molecule

[0038] Among them, E at is the atomization energy of the sample configuration, E sum_atoms is the sum of the energies of the atoms that make up the molecule, E molecule is the total energy of the molecule.

[0039] Step S300, based on the atomization energy and its force data calculated by density functional theory, and the atomization energy and its force data of the corresponding sample configuration calculated by density functional tight binding theory, calculate the repulsion energy and repulsion force of each sample configuration.

[0040] In this embodiment, the atomization energy and force parameters obtained by density functional theory (DFT) calculation are first obtained; then, the repulsion energy and repulsion force are calculated based on these parameters and the atomization energy and force data of the corresponding sample configuration calculated by density functional tight binding (DFTB).

[0041] In this embodiment, for the same configuration, the density functional tight binding-neural network hybrid potential function constructed by the present invention and the atomization energy and the corresponding force calculated by density functional theory should be equal, that is, E at_DFTB_NN =E at_DFT ,Therefore, the calculation process of repulsion energy and its force is:

[0042] E rep =E at_DFTB_el -E at_DFT

[0043] F rep =F DFTB_el -F DFT

[0044] Among them, E at_DFTB_el The band structure atomization energy, E, obtained by density functional tight binding calculation for each sample configuration at_DFT is the atomization energy calculated by density functional theory for each sample configuration, F DFTB_el The density functional tight binding force data for each sample configuration, F DFT Density functional theory force data for each sample configuration.

[0045] Step S400, constructing an atomic neural network potential function based on the above calculation results, and using repulsive energy and repulsive force as labels to train the neural network potential function.

[0046] Optionally, the use of repulsive energy and repulsive force as labels to train the neural network potential function includes: obtaining a density functional tight binding data set based on density functional tight binding calculation results; dividing the density functional tight binding data set into a training set, a validation set and a test set; using the repulsive energy and repulsive force in the training set as labels to train the neural network potential function, the validation set is used to adjust the samples, and the test set is used to evaluate the performance of the trained neural network potential function.

[0047] Optionally, after dividing the density functional tight binding data set into a training set, a validation set and a test set, it also includes: creating a data loader, reading the training set, the validation set and the test set into the memory of the data loader, and performing preprocessing to obtain a data form that can be used by the neural network potential function.

[0048] In this embodiment, data loaders are created for the training set, validation set, and test set respectively, a SchNet model is constructed, and corresponding hyperparameters are set. The constructed data loader reads the original data from the corresponding data set, and then converts the format of the read data to obtain a data form suitable for use by the neural network potential function.

[0049] Optionally, an atomic neural network potential function is constructed based on the above calculation results. Neural network potential function fitting software including but not limited to SchNetPack, DeePMD-Kit, LASP, etc. can be used to fit the multi-body local interactions in the density functional tight binding data set to obtain the neural network potential function.

[0050] In this embodiment, the data set obtained by the above calculation process is used to train the atomic neural network, and the E rep and F rep As labels, the atomic neural network model is trained. The multi-body local interaction in DFTB is fitted using neural network potential function fitting software, such as SchNetPack, DeePMD-Kit, LASP, etc., to obtain the atomic neural network potential function that describes the repulsive energy of a specific system.

[0051] Step S500 , combining the trained neural network potential function with density functional tight binding theory to generate a hybrid potential function, performing error analysis on the hybrid potential function, and optimizing the hybrid potential function according to the analysis result.

[0052] In this embodiment, the repulsive energy and repulsive force are described based on the neural network potential function; the density functional tight binding band structure energy and the neural network repulsive energy are combined, and the potential function force and the neural network repulsive force are combined; and the hybrid potential function is subjected to an error analysis based on the combination result.

[0053] In this embodiment, the trained neural network potential function is combined with the DFTB method to generate a hybrid potential function. The E calculated by the DFTB+ software is summed using the ase.calculators.mixing.SumCalculator class in the Atomic Simulation Environment (ASE) software. DFTB_el and F DFTB_el The result E obtained with the NN potential function NN_rep and F NN_rep Combine and derive the DFTB-NN rep The potential function is used for error analysis and optimization, and finally an accurate and efficient DFTB-NN is obtained. rep Potential function.

[0054] The method for constructing a density functional tight binding-neural network hybrid potential function provided in this embodiment uses a machine learning method to perform multi-dimensional fitting of the repulsive energy in DFTB to generate a high-precision DFTB-NN hybrid potential function model. Since only the band structure energy part in DFTB needs to be calculated in this method, its calculation efficiency has been greatly improved compared with the SCC-DFTB method. The repulsive part is fitted by a neural network, which makes up for the deficiency of the atomic pair potential in the traditional DFTB method in describing multi-dimensional problems in complex structures, and significantly improves the prediction accuracy. Compared with the use of the DFTB method or the traditional approximate model alone, the present invention can more accurately predict the energy and force of the target material structure in the process of constructing the hybrid potential function. In the present invention, the energy and force of the system are decomposed, and the DFTB method based on the physical model is used to calculate part of it, and the repulsive part is trained by the neural network, which reduces the dependence of machine learning on a specific data set, thereby enhancing the transferability of the model, so that the method of the present invention is not only applicable to specific material systems, but also can be extended to other material systems. By using the combination of the DFTB method and the neural network, the present invention significantly reduces the calculation time and resource consumption. In practical applications, this means that the calculation and analysis of a large number of materials can be completed in a shorter time, greatly improving the research efficiency. The calculation methods and software tools used in the present invention, such as DFTB+ software, SchNetPack, DeePMD-Kit, LASP and ASE, are all mature and widely used software packages in the prior art. Researchers can easily integrate the DFTB-NN hybrid potential function construction method of the present invention into the existing calculation process without large-scale software and hardware transformation.

[0055] like Figure 2 As shown, the present invention takes a small molecule system as an example to construct a DFTB-NN hybridization potential function of a small molecule system {C, H, O, N, F} whose target material is a small molecule system.

[0056] In this example, the DFTB-NN hybrid function was constructed for the QM9 data set of the small molecule system. The DFTB+ software was combined with the DFTB 3ob parameter set, and the neural network fitting was performed using the SchNetPack software package. The construction process is as follows: Figure 2 shown.

[0057] In this embodiment, the QM9 data set is constructed based on density functional theory. The QM9 data set contains detailed information on 133,885 small molecules, which are composed of five elements: carbon (C), hydrogen (H), oxygen (O), nitrogen (N) and fluorine (F), and each molecule contains up to 9 heavy atoms (CONF). The 3D structure of each molecule in the QM9 data set includes the spatial coordinates of all atoms, as well as energy and force data calculated using the density functional theory (DFT) method (i.e., this embodiment uses the B3LYP functional and the 6-31G (2df, p) basis set).

[0058] In this embodiment, the band structure energy E corresponding to the electronic structure of each target material sample in the QM9 data set is calculated using the DFTB+ software combined with the 3ob parameter set. DFTB_el and force F DFTB_el , excluding the repulsive term, the result E DFTB_el and F DFTB_el Write to the database (.db) file of ase.

[0059] The single-atom energy of CHONF is obtained by DFTB+, and the repulsive energy Erep and the force Frep are calculated by the following formula:

[0060] E at =E sum_atoms -E molecule

[0061] E rep =E at_DFTB_el -E at_DFT

[0062] F rep =F DFTB_el -F DFT

[0063] The rejection energy E rep And force F rep Written into the database (.db) file of ase for subsequent neural network potential function training.

[0064] Use the data set obtained in the above steps to train the neural network and use E in the data set rep and Frep As labels, train the neural network model. Divide the dataset into training set, validation set and test set, create data loaders for each set, build the SchNet model, and set the corresponding hyperparameters. The atomic basis dimension and the number of filters are set to 30, 20 Gaussian functions, 5 interaction layers, and the cutoff distance is The truncated network is Cosine Cutoff, and the learning rate is updated using ReduceLROnPlateau (patience value is 5, factor is 0.8, minimum learning rate is 1e-6), the initial learning rate is 0.01, GPU is used for training, and the Adam optimizer is used.

[0065] In this embodiment, in order to demonstrate the advantages of the constructed DFTB-NN potential function, two training strategies were used for training. Strategy one is a training set of 1000 items, a validation set of 500 items, and the rest are test sets (132,385 items); Strategy two is a training set of 10,000 items, a validation set of 5,000 items, and the rest are test sets (118,885 items). By comparing the results of these two strategies, the performance improvement of the DFTB-NN hybrid potential function in small molecule systems is verified. Under the same training conditions, the accuracy of the DFTB-NN potential function in the validation set and test set of the QM9 data is at least 1 times higher than that of the NN potential function, showing that the DFTB-NN hybrid potential function has better transferability, which is specifically manifested in that it can better predict the energy and force of species in the validation set and test set by training in a smaller training set. The test results are as follows: Figure 3 shown.

[0066] In this embodiment, the trained neural network potential function is combined with the DFTB method, and the ase.calculators.mixing.SumCalculator class in the AtomicSimulation Environment (ASE) software is used to set the E calculated by the DFTB+ software. DFTB_el and F DFTB_el The NN potential function calculation result E obtained by Schnet training NN_rep and F NN_rep By combining them, the DFTB-NN potential function can be used in ASE.

[0067] The present invention also takes the iron-based catalytic system as an example to construct a DFTB-NN hybrid potential function in which the target material is the catalytic system {Fe, C, H, O}.

[0068] In this example, a DFTB-NN hybrid function was constructed for a data set of an iron-based catalytic system, using DFTB+ software combined with the trans3d parameter set of DFTB, and the neural network fitting was performed using the LASP software package.

[0069] In this embodiment, the data set of the iron-based catalytic system includes 21,412 structural data composed of four elements: Fe, C, H, and O, and energy stress data obtained by DFT calculation, including bulk, surface, and cluster structures.

[0070] In this example, the DFTB+ software was used to calculate the band structure energy E corresponding to the electronic structure configuration of each sample in the FeCHO data set in combination with the trans3d parameter set. DFTB_el and force F DFTB_el , excluding the repulsive term, the result E DFTB_el and F DFTB_el Write to the database (.db) file of ase.

[0071] In this embodiment, the single atom energy of FeCHO is obtained by DFTB+, and the repulsion energy E is calculated by the following formula: rep And force F rep get:

[0072] E at =E sum_atoms -E molecule

[0073] E rep =E at_DFTB_el -E at_DFT

[0074] F rep =F DFTB_el -F DFT

[0075] The rejection energy E rep And force F rep Written into the database (.db) file of ase for subsequent neural network potential function training.

[0076] Use the data set obtained in the above steps to train the neural network and use E in the data set rep and F rep As a label, convert .db into the training set input file format required by LASP software, and use LASP software to train the neural network model. Two training strategies are used for training:

[0077] (a) Loss function E:F:Stress = 200:1000:10, the number of neurons in the input layer is 305, the number of neurons in the hidden layer is 80 / 50 / 50, and the number of neurons in the output layer is 1; the number of NN potential function parameters is 124524. (b) Loss function E:F:Stress = 200:15:1, the number of neurons in the input layer is 543, the number of neurons in the hidden layer is 112 / 80 / 80, and the number of neurons in the output layer is 7; the number of NN potential function parameters is 310076. Both are trained for 20,000 epochs, and the test results are shown in Figure 2. Figure 3 As shown. NN_Liu is the accuracy of the NN potential function in the literature, NN_a and NN_b are the NN potential functions obtained by testing and training in this embodiment, and DFTB_a and DFTB-NN_b are the DFTB-NN hybrid potential functions in this example. Figure 4 As shown in the figure, compared with the training results under the same conditions, in strategy a with fewer NN potential function parameters, the results of DFTB-NN_a are significantly improved compared with NN_a. In strategy b with more NN potential function parameters, the potential functions of DFTB-NN_b and NN_b are both highly accurate and meet the usage requirements.

[0078] In this embodiment, the DFTB-NN potential function constructed in this embodiment has greatly improved accuracy and computational efficiency compared with the traditional DFTB method. Figure 5 As shown, the error is significantly reduced in the test of FeC cluster structures outside the 3500 training sets.

[0079] In this embodiment, the trained neural network potential function is combined with the DFTB method, and the ase.calculators.mixing.SumCalculator class in the AtomicSimulation Environment (ASE) software is used to set the E calculated by the DFTB+ software. DFTB_el and F DFTB_el The NN potential function calculation result E obtained by LASP training NN_rep and F NN_rep By combining them, the DFTB-NN potential function can be used in ASE.

[0080] As one or more specific application embodiments of the embodiments of the present invention, the optimal implementation scheme or the scheme that the inventor most wants to embody is described in combination with specific application scenarios.

[0081] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A method for constructing a density functional tight binding-neural network hybrid potential function, characterized in that: The method comprises: Construct a density functional theory data set, obtain the structural parameters of the target material samples in the data set, and calculate the energy, atomization energy and force data of the electronic structure corresponding to the sample configuration based on density functional theory; Calculating the band structure energy and the force of the target material sample in the density functional theory data set based on the density functional tight binding theory, and obtaining the corresponding density functional tight binding atomization energy; Based on the atomization energy and its force data calculated by density functional theory, and the atomization energy and its force data of the corresponding sample configurations calculated by density functional tight binding calculation, the repulsion energy and repulsion force of each sample configuration are calculated; Constructing an atomic neural network potential function based on the above calculation results, and using repulsive energy and repulsive force as labels to train the neural network potential function; The trained neural network potential function is combined with the density functional tight binding theory to generate a hybrid potential function, and an error analysis is performed on the hybrid potential function, and the hybrid potential function is optimized according to the analysis result.

2. The method for constructing a density functional tight binding-neural network hybrid potential function according to claim 1, characterized in that: The energy, atomization energy and force data of the electronic structure corresponding to the sample configuration calculated based on density functional theory also include: Obtain the sum of the energy of each atom and the total energy of the molecule that constitutes the target material sample; The atomization energy of each sample configuration is calculated based on the sum of the energies of the atoms that make up the molecule and the total energy of the molecule.

3. The method for constructing a density functional tight binding-neural network hybrid potential function according to claim 2, characterized in that: The formula for calculating the atomization energy of each sample configuration is: AND at =And sum_atoms -AND molecule Among them, E at is the atomization energy of the sample configuration, E sum_atoms is the sum of the energies of the atoms that make up the molecule, E molecule is the total energy of the molecule.

4. The method for constructing a density functional tight binding-neural network hybrid potential function according to claim 1, characterized in that: The calculation formula for calculating the repulsive energy and repulsive force of each sample configuration is: AND rep =And at_DFTB_el -AND at_DFT F rep =F DFTB_el -F DFT Among them, E at_DFTB_el The band structure atomization energy, E, obtained by density functional tight binding calculation for each sample configuration at_DFT is the atomization energy calculated by density functional theory for each sample configuration, F DFTB_el The density functional tight binding force data for each sample configuration, F DFT Density functional theory force data for each sample configuration.

5. The method for constructing a density functional tight binding-neural network hybrid potential function according to claim 1, characterized in that: The error analysis of the hybrid potential function is performed as follows: Describing repulsive energy and repulsive force based on the neural network potential function; Combining the density functional tight binding band structure energy and the neural network repulsion energy, and combining the potential function force and the neural network repulsion force; An error analysis is performed on the hybrid potential function according to the combination result.

6. The method for constructing a density functional tight binding-neural network hybrid potential function according to claim 1, characterized in that: The method of using the repulsive energy and the repulsive force as labels to train the neural network potential function includes: Obtain a density functional tight binding data set based on density functional tight binding calculation results; Dividing the density functional tight binding data set into a training set, a validation set and a test set; The repulsive energy and repulsive force in the training set are used as labels to train the neural network potential function, the validation set is used to adjust the samples, and the test set is used to evaluate the performance of the trained neural network potential function.

7. The method for constructing a density functional tight binding-neural network hybrid potential function according to claim 6, characterized in that: After dividing the density functional tight binding data set into a training set, a validation set and a test set, it also includes: A data loader is created, the training set, validation set and test set are read into the memory of the data loader, and preprocessed to obtain a data form that can be used by the neural network potential function.

8. The method for constructing a density functional tight binding-neural network hybrid potential function according to claim 1, characterized in that: Based on the above calculation results, the atomic neural network potential function is constructed. The neural network potential function fitting software including but not limited to SchNetPack, DeePMD-Kit, LASP, etc. can be used to fit the multi-body local interactions in the density functional tight binding data set to obtain the neural network potential function.

Images